Artificial intelligence manages to create images from words
Nourished by hundreds of millions of images gleaned from the web, the Dall.E neural network is very creative
The company OpenAI, specializing in artificial intelligence , presented Tuesday, January 5 a tool to generate images from a simple text, reports 01Net . Called Dall.E, it is a neural network that is based on GPT-3, the language model developed by the company founded in 2015 by Elon Musk .
Artificial intelligence is progressing well. Here is dall.E an AI from openAI that generates images from descriptive text. Frankly the lawyer chairs we are clearly in the "artistic creation". More information: https://t.co/3Vt82ZdB6j pic.twitter.com/REmcnHrxWs- Thomas Cabaret (@ ThomasCabaret84) January 8, 2021
The name Dall.E is the contraction of “Dali”, the famous Spanish surrealist painter, and “Wall-E”, the cute robot from Pixar studios. A particularly apt name, given the creative faculties of the system. For example, OpenAI researchers provided Dall.E with the words “chair” and “lawyer”. The AI then combined the two concepts to provide different convincing lawyer chair proposals.
Billions of images
OpenAI took it a step further by providing the AI with the labels “baby radish in tutu” and “take a dog for a walk”. Dall.E made many proposals for images, being very original. To achieve these kinds of results, the software was trained with twelve billion captioned images found on the web .
The company also presented Clip (Contrastive Language-Image Pre-training), another neural network that manages to do the opposite of Dall.E: recognize categories of objects when images are submitted to it.